auto-ontology: Ground Agent Queries in Enterprise Data with NVIDIA NIM

Overview

NVIDIA Auto Ontology is a semantic, graph-based context layer for agents working with enterprise data. Built around NVIDIA NIM for inference, it turns metadata from connected data sources into an ontology that helps agents understand what metrics and entities mean, where they come from, and how they relate. The aim is to make agentic analytics more reliable than querying databases without business context.

It complements existing semantic models and makes the resulting context available to agent harnesses through MCP, a CLI, and an API. The project includes a web application and supports local deployment with Docker Compose or deployment to Kubernetes.

Key Features

• Generates an ontology from enterprise metadata and connects existing ontologies and semantic models.
• Reconciles differing definitions of business metrics and entities across systems.
• Translates natural-language questions into SQL and executes queries against connected databases.
• Checks generated queries against certified definitions and lineage, with provenance returned alongside answers.
• Provides a web UI for chat, analysis, catalog management, analytics, and settings.
• Offers authenticated API access, including user-scoped API tokens and streaming chat responses.
• Exposes data context to agent harnesses through an MCP server.
• Supports deployments with Docker Compose or a Kubernetes Helm chart.

Use Cases

• Data and analytics teams can let business users ask questions about metrics in natural language while grounding answers in catalog definitions.
• Agent developers can connect MCP-compatible tools or applications to enterprise data with a shared semantic context layer.
• Data stewards can reconcile definitions across sources and check query results against certified definitions and lineage.
• Platform teams can deploy a self-hosted analytics application for connected databases, using Docker Compose for a local stack or Helm for Kubernetes.

Getting Started

For a local Docker Compose setup, clone the repository, configure .env from .env.example with database and model credentials, then run:

docker compose up -d --build

See the README and deployment guide for configuration and Kubernetes instructions.

Alternatives

• QueryWeaver: QueryWeaver generates SQL from schema graphs for PostgreSQL and MySQL, while auto-ontology builds a broader semantic context layer for agents.
• vanna: Vanna focuses on database chat interfaces that generate SQL and return results, while auto-ontology provides reusable enterprise context through MCP, a CLI, and an API.

Considerations

• The local stack requires Docker and Docker Compose, plus Postgres and an NVIDIA API key for the default NVIDIA NIM setup. Kubernetes deployment is also supported; the documented Python runtime is 3.12 or 3.13.
• UI-added database credentials are stored in plaintext in Postgres unless Vault is fully configured. Review the connection-management guidance and configure Vault if plaintext credential storage is unsuitable.
• The repository is active, with regular recent commits and a dozen contributors, but it is still young and has no release in the past year. The measured contributor data does not indicate dependence on one person.
• The project reports no test suite in the supplied project facts. It is also not accepting external contributions, which may affect teams that expect to contribute fixes upstream.
• Apache-2.0 applies to the project; third-party components have separate notices. Runtime model APIs and connected data systems are subject to their own terms.